[Model] Add DeepSeek-V4 CPU backend (#55355)

Signed-off-by: jiang1.li <jiang1.li@intel.com>
This commit is contained in:
Li, Jiang
2026-09-10 14:14:48 +08:00
committed by GitHub
parent 49632fb3ac
commit c3ccc0e957
47 changed files with 9564 additions and 627 deletions
@@ -70,12 +70,14 @@ def main(
packed_w13 = cpu_prepack_moe_weight(w13, isa)
packed_w2 = cpu_prepack_moe_weight(w2, isa)
output = torch.empty_like(input_tensor)
def run_benchmark(iters: int) -> list[float]:
times = []
for _ in range(iters):
start_time = time.perf_counter_ns()
_ = cpu_fused_moe(
cpu_fused_moe(
output,
input_tensor,
packed_w13,
packed_w2,